【作者】 王建华；

【导师】 王中秋；

【作者基本信息】 南京中医药大学 ， 中西医结合临床（专业学位）， 2019， 博士

【摘要】 第一部分理论研究胰腺实性-假乳头状瘤(Solid Pseudopapilla ry Neoplasm of the pancreas,SPN)是一种胰腺少见低度恶性肿瘤,但近年来临床检出率明显升高。目前手术切除是治疗胰腺SPN的主要方法,也是唯一治愈途径。传统的胰腺恶性肿瘤根治术不仅创伤大、风险高,而且易造成患者术后胰腺内、外分泌功能不全,加上大多数胰腺SPN具有惰性的生物学行为,因此外科医生目前倾向于对胰腺SPN采取较小的手术方式,但由于肿瘤存在异质性,约14.4%～33.3%的肿瘤具有侵袭性,可侵犯邻近组织或出现远处转移,而较小的手术方式并不一定能完整切除肿瘤组织,容易造成肿瘤术后残留、复发。因此术前准确判断肿瘤的侵袭性是影响手术方案制定的关键因素。胰腺SPN多发生于年轻女性,社会危害性大,并且肿瘤缺乏特异性的临床症状与实验室标记物,加上穿刺活检并不能全面、准确地反映肿瘤的异质性。由于目前CT检查是临床诊断胰腺SPN的主要手段,而对于CT征象在侵袭性与非侵袭性胰腺SPN中的鉴别价值尚存在争议,因此探讨术前CT征象对胰腺SPN侵袭性的预测效能具有重要的临床价值。近年来随着精准医疗、个体化治疗的快速发展,影像组学已成为当前临床研究的热点,它主要通过机器学习的方法从大量的医学图像(CT、MRI或PET)中提取的高通量定量特征,并将其转换成可采集的高维数据。影像组学本质是一种无创的定量分析方法,与穿刺活检通过组织病理学评估肿瘤异质性不同,影像组学通过描绘图像中的不均质性,从而反映肿瘤生长的微环境和肿瘤本身的异质性,因此具有无创性、经济性和可重复性,并且不会给患者带来额外的负担与风险,而且经分析获得的定量数据可以使临床决策变得更为稳定一致。目前影像组学研究已在临床广泛开展,涵盖了疾病诊断与鉴别诊断、生物学行为评估、病理分型和分级、肿瘤分期与疗效预测、疾病预后及生存预测等诸多领域,并且都展现出了较高的临床价值。第二部分 基于CT征象对胰腺实性-假乳头状瘤侵袭性的预测研究研究目的:基于病理金标准,通过分析侵袭性和非侵袭性胰腺SPN的多期增强CT表现,探讨常规MSCT上影像特征在预测SPN侵袭性中的临床价值,并且通过联合临床特征构建胰腺SPN侵袭性的预测模型。材料和方法:回顾性分析经手术病理证实为胰腺SPN的127例患者的临床及影像资料,根据术后病理表现分为侵袭性和非侵袭性胰腺SPN,分别对两组人群的年龄、性别、临床症状、手术方式、肿瘤CT征象(位置、大小、形态、质地、边界、外生性、包膜、出血、钙化、肿瘤最大径、CT值、CT 比值、强化方式)及肿瘤周围的间接征象进行比较。使用独立样本t检验或Mann-Whitney U检验来比较两组间连续型变量的差异,采用x2检验或Fisher确切概率检验比较两组间定性变量的差异,然后将单因素分析中具有统计学差异的参数纳入到多因素logistic回归中,采用逐步后退的LR法筛选出预测胰腺SPN侵袭性的独立危险因素,并构建相应预测模型,再运用ROC曲线计算模型的预测效能。结果:最终纳入127例胰腺SPN患者,其中侵袭性组有32例,非侵袭性组有95例。单因素分析发现年龄、性别、肿瘤最大径、包膜、边界及肿瘤内出血在两组间存在统计学差异(p值分别为0.017、0.022、0.006、<0.001、0.026、0.017)。侵袭性组与非侵袭性胰腺SPN组相比,患者平均年龄增大,男性患者比例增高,肿瘤中位最大径减小,肿瘤缺乏完整包膜及边界不清的比例增高,而肿瘤内出血的比例降低。在常规CT特征的多因素logistic回归中,发现肿瘤缺乏完整包膜是预测胰腺SPN侵袭性的独立危险因素[OR值为6.259,96%CI(2.221～17.637),p=0.001],此时模型AUC值为0.69,95%CI为0.59～0.79,敏感性为84.4%、特异性为53.7%、正确性为61.4%;在联合预测模型中肿瘤缺乏完整包膜仍然是预测胰腺SPN侵袭性的独立危险因素,模型的AUC值为0.739,95%CI为0.639～0.84,敏感性为81.3%、特异性为57.9%,正确性为76.4%。结论:1.患者性别、年龄、肿瘤最大径、包膜、边界及肿瘤内出血在侵袭性与非侵袭性胰腺SPN中存在统计学差异。2.肿瘤缺少完整包膜是预测胰腺SPN侵袭性的独立危险因素。3.CT特征与患者年龄的联合模型对预测胰腺SPN的侵袭性具有一定价值。第三部分 基于影像组学模型对胰腺实性-假乳头状瘤侵袭性的预测研究研究目的:由于影像组学可以定量评估肿瘤的异质性,本研究基于术前常规CT图像,通过影像组学与机器学习的方法,构建预测胰腺SPN侵袭性的影像组学标签及列线图模型,从而探讨影像组学在预测胰腺SPN侵袭性中的临床价值。材料和方法:回顾性收集从2012年10月至2018年10月间经手术病理证实的112例胰腺SPN患者的临床及影像资料,以7:3的比例,将2012年10月至2016年6月间的77例作为训练组,将2016年7月至2018年10月间的35例作为验证组,根据术后病理分为侵袭性和非侵袭性胰腺SPN。分别对训练组、验证组中侵袭性与非侵袭性胰腺SPN患者的临床资料及常规CT征象进行比较,将训练组中存在统计学差异的变量纳入多因素logistic回归,通过逐步后退的LR法筛选出预测胰腺SPN侵袭性的独立危险因素,并构建预测模型。在影像组学分析上,首先使用AK软件对所有患者的三期CT图像进行预处理,然后使用ITK-SNAP软件对三期图像分别进行ROI分割,再运用AK软件提取影像组学特征。按照扫描图像的不同组合分别建立基于平扫、胰实质期、门静脉期、平扫联合胰实质期、平扫联合门静脉期、以及平扫联合胰实质与门静脉期图像的6种影像组学标签。在影像组学特征筛选中,首先进行组内、组间相关系数的一致性检验(ICC),以评估图像分割的可重复性及提取的影像组学特征的稳定性,对稳定性高的影像组学特征参数进行数据预处理,然后再分别通过单因素分析、相关性分析和含十折交叉验证的LASSO回归对6种影像组学特征参数进行降维筛选,从而构建影像组学标签,再运用ROC曲线评价每种影像组学标签在训练组、验证组及全部112例患者中的预测效能,最终选取在验证组中预测性能最佳的影像组学标签联合患者年龄与常规CT特征进行多因素logistic回归分析,并构建列线图模型。运用ROC曲线、Bootstrap自助抽样及Hosmer-Lemeshoe拟合优度检验评价列线图模型在训练组与验证组中的预测效能,使用Delong检验比较各模型间的预测效能是否具有显著差异,采用列线图将模型的预测结果可视化,使用校正曲线、决策曲线评价列线图的校正效果及临床价值。结果:本组中的112例患者与上一部分的127例患者在临床资料与常规CT征象上均无统计学差异,并且本组中侵袭性与非侵袭性胰腺SPN间存在统计学差异的变量与上一部分一致。本组训练组与验证组中除肿瘤质地外,患者临床资料、常规CT征象均无明显统计学差异。影像组学特征的组内、组间ICC均大于0.75,因此将396个影像组学特征全部纳入到特征降维筛选中,发现基于胰实质期图像的影像组学标签在验证组中的预测效能最高,AUC值为0.88,因此列线图预测模型由患者年龄、常规CT特征与胰实质期的影像组学标签构成,经多参数logistic逐步回归筛选,发现影像组学标签与肿瘤包膜是预测胰腺SPN侵袭性的独立危险因素,肿瘤包膜的调整OR值为7.635,95%CI为1.709～34.108,P值为0.008;影像组学标签的调整OR值为9.849,95%CI为2.314～41.912,P值为0.002。列线图模型在训练组中预测效能的AUC值为0.856。Delong检验发现在训练组中,列线图模型的预测效能高于临床信息模型、常规CT特征模型及联合模型,且差异具有统计学意义(p值分别为0.0177、0.0006、0.0006)。在验证组中,列线图模型预测效能的AUC值为0.932。Hosmer-Lemeshoe拟合优度检验在训练组和验证组间差异均无统计学意义(p值分别为0.794、0.668)。决策曲线分析显示列线图模型具有较高的临床价值。结论:1.6种影像组学标签对胰腺SPN的侵袭性均有较好的预测效能,其中基于胰实质期图像构建的影像组学标签在验证组和全部112例胰腺SPN中预测效能最高。2.在训练组的常规CT特征模型及联合模型中,肿瘤缺乏完整包膜是预测胰腺SPN侵袭性的独立危险因素;在列线图模型中,影像组学标签和肿瘤缺乏完整包膜是预测胰腺SPN侵袭性的独立危险因素。3.在训练组多个模型的性能比较中,列线图模型的预测效能高于临床资料、常规CT特征以及联合模型,且Delong检验表明差异具有统计学意义。列线图模型在训练组与验证组中均具有较高的预测效能,且Hosmer-Lemeshoe拟合优度检验表明模型在训练与验证组中没有偏离拟合,决策曲线分析显示列线图模型具有较高的临床价值。更多还原

【Abstract】 PartⅠ:the theoretical studySolid Pseudopapillary Neoplasm(SPN)is a rare and low-grade malignant tumor of the pancreas,but its clinical incidence has increased significantly in recent years.At present,surgical resection is the main method for SPN.Traditional pancreatic malignant radical surgery is not only traumatic,high risk,but also easy to cause postoperative pancreatic endocrine and exocrine insufficiency,and most of the pancreatic SPN inert biological behavior,so surgeons tend to take minimally invasive surgery.However,due to the heterogeneity between SPNs,about 14.4%～33.3%of the tumors are invasive and can invade adjacent tissues or distant metastases.The minimally invasive surgery may not completely remove the tumor tissue,which may cause recurrence and metastasis.Therefore,accurately determining the invasiveness of the tumor before surgery is a key factor in determining the surgical plan.SPN occurs mostly in young women,and it is socially harmful.The lack of specific clinical symptoms and laboratory markers in the tumor,needle biopsy cannot fully and accurately reflect the heterogeneity of the tumor.CT is the main means in clinical diagnosis,but the current value of CT signs in the invasive and non-invasive pancreatic SPN is still controversial.Therefore,it is of great clinical value to explore the preoperative CT signs to predict the invasiveness of SPN.In recent years,with the rapid development of precision medicine and individualized treatment,radiomics has become a hot spot in current clinical research.It mainly extracts high-throughput quantitative features from a large number of medical image images(CT,MRI or PET)by machine learning and converts them into high-dimensional data that can be collected.The essence of radiomics is a non-invasive quantitative analysis,which is different from the biopsy to assess the heterogeneity of tumors by histopathology.Radiomics reflects the microenvironment of tumor growth and the tumor itself by depicting the heterogeneity in the image.Heterogeneity,non-invasive,economical,reproducible,and does not impose additional burdens and risks on patients,and the quantitative data obtained from the analysis can make clinical decisions more stable and consistent,so radiomics has been widely carried out in clinical practice,covering diseases diagnosis and differential diagnosis,biological behavior assessment,pathological and grading.tumor staging,efficacy prediction and evaluation,disease prognosis and survival prediction,and all of them show high clinical ValuePartⅡ:Prediction of invasion of pancreatic solid-pseudopapillary tumor using CT findingsPurposes:Based on the pathological gold criteria,the multi-phase enhanced CT findings of invasive and non-invasive pancreatic SPN were analyzed to investigate the diagnostic value of conventional MSCT imaging features in predicting SPN invasiveness.and a predictive model of SPN invasiveness was constructed by combining clinical features.Materials and Methods:The clinical and imaging data of 127 patients with pancreatic SPN confirmed by surgical pathology were retrospectively analyzed.According to the postoperative pathology,they were divided into invasive pancreatic SPN or non-invasive pancreatic SPN.The age,sex and clinical symptoms,surgical methods,imaging signs(location,size,morphology.texture,boundary,exogenous,envelope,hemorrhage,calcification,pancreaticobiliary dilatation,maximum tumor diameter,CT value,CT ratio,enhancement model)and surrounding tumors indirect signs of the two groups were respectively compared.The difference of continuous variables between the two groups was compared with the independent sample t test or Mann-Whitney U test,and the difference of qualitative variables between the two groups was compared by x2 test or Fisher exact test method.Then the parameters of single factor analysis with statistical difference were incorporated into binary logistic regression analysis,and the independent risk factors for predicting the invasion of pancreatic SPN were screened by stepwise backward LR,and the predictive model of combined CT features and clinical information was constructed.The diagnostic efficiency of the model was calculated by ROC curve.Results:Finally,127 patients with pancreatic SPN were included,including 32 in the invasive group and 95 in the non-invasive group.Univariate analysis revealed that patients’age,gender,tumor maximum diameter,capsule,border,and intratumoral hemorrhage were statistically different between the t,^wo group(p=0.017,0.022,0.006,<0.001,0.026,0.017,respectively).Compared with the non-invasive pancreatic SPN group,the average age has increased,the proportion of male patients has increased,and the median diameter of the tumor has decreased,and the proportion of tumor lacking intact capsule and unclear border has increased,and the proportion of bleeding has decreased in the invasive pancreatic SPN group.After the multi-factor logistic regression of conventional CT features with stepwise,it was found that the lack of complete capsule of tumor was an independent risk factor to predict the invasion of SPN[OR=6.259,96%CI(2.221～17.637),p=0.001],The model AUC value was 0.69,95%CI was 0.59～0.79,sensitivity was 84.4%,specificity was 53.7%,and correctness was 61.4%.The combined model had an AUC value of 0.739 and 95%CI of 0.639～0.84.The sensitivity was 81.3%,the specificity was 57.9%,and the correctness was 76.4%.Conclusions:1.Patients’ gender,age,tumor maximum diameter,capsule,border,and intratumoral hemorrhage were statistically different between invasive and non-invasive pancreatic SPN.2.The lack of a complete capsule of the tumor is an independent risk factor for predicting the invasiveness of SPN.3.The combined model of conventional CT image features and t age has certain value for predicting the invasiveness of pancreatic SPN.Part Ⅲ:CT radiomics for predicting the invasiveness of pancreatic SPNPurposes:Because radiomics can quantitatively assess the heterogeneity of tumors,this study based on preoperative routine CT images through radiomics and machine learning methods to construct an radiomics label and nomogram model for predicting the invasiveness of pancreatic SPN,and to explore the clinical value of radiomics in the prediction aggressive of pancreatic SPN.Materials and Methods:The clinical and imaging data of 112 patients with pancreatic SPN confirmed by surgery and pathology from October 2012 to October 2018 were retrospectively collected.77 patients were used as a training group by the ratio of 7:3 between October 2012 and June 2016,35 patients bet.ween July 2016 and October 2018 were used as a validation group,and all patients were divided into invasive or non-invasive pancreatic SPN according to postoperative pathology.The clinical data and routine CT signs of patients with invasive and non-invasive pancreas in the training and validation groups were compared.The statistically significant variables in the training group were included in the multivariate logistic regression,and the LR method was used to screen through the stepwise retreat.Predicting independent risk factors for pancreatic SPN invasiveness and constructing predictive models.In the radiomics analysis,the three-phase CT images of all patients were pre-processed,and then the thr-ee-phase images were segmented layer by layer using ITK-SNAP,and then the AK was used to extract the radiomics f-eatures.Six radiomics labels based on the images of plain scan,pancreatic parenchyma,portal vein,plain scan combined with pancreatic parenchyma,plain scan combined with portal vein and plain scan combined with pancreatic parenchyma and portal vein were established according to different scanned images.Before the screening of radiomics features,the intra-and inter-group correlation coefficients was used to evaluate the reproducibility of image segmentation and the stability of photographic genre features.The data was preprocessed,and then the phenotypic parameters of the six imaging labels were screened by single factor analysis,correlation analysis and LASSO regression with ten-fold cross-validation,and the radiomics labels were calculated and evaluated by ROC.The predictive efficacy of each radiomics label in the training,validation,and all 112 patients were compared,and finally selected the best predictive radiomics label in the validation group,multivariate logistic regression analysis by combining patient age with conventional CT f-eatures and build a nomogram model.ROC curve,AUC value,Bootstrap and HL test were used to evaluate the predictive performance of the nomogram model in the training and validation groups.Delong test was used to compare whether the prediction performance between the models was significantly different.The prediction results of the model are visualized by using the nomogram.and the correction effect and clinical value of the nomogram are evaluated using the calibration curve and the decision curveResults:There were no statistically significant differences in clinical data and routine CT findings between 112 patients in this group and 127 patients in the previous section.There were no statistically significant differences in the clinical data,routine CT signs,and invasive patients between the training group and the validation group.The intra-and inter-group ICC of tumor segmentation was greater than 0.75,so all 396%radiomics features were included in the feature screening.After date analysis,it was found that the radiomics label based on the pancreatic parenchymal phase image had the highest predictive power in the validation group,and the AUC value was 0.88.So the nomogram prediction model consists of patient age,conventional CT features and radiomics label based on the pancreatic parenchymal phase image s.After multi-parameter logistic regression,the radiomics label and tumor capsule are found as an independent risk factor for predicting the invasiveness of pancreatic SPN.The adjusted OR value of tumor capsule was 7.635,95%CI was 1.709 to 34.108,P value was 0.008,and the adjusted OR value of radiomics label was 9.849,95%CI was 2.314～41.912,P value is 0.002.The AUC value of the predictive performance of the nomogram model in the training group is 0.856.Delong test found that the predictive power of the nomogram model is higher than the clinical information model,the conventional CT feature model and the combine model,and the difference is statistically significant(p=0.0177,0.0006,0.0006,respectively).In the validation group,the AUC value of the prediction performance of the nomogram model is 0.932.The Hosmer-Lemeshoe goodness-of-fit test was not statistically significant between the training and validation groups(p=0.794,0.668,respectively).Decision curve analysis shows that the nomogram model has high clinical value.Conclusions:1.Six radiomics label have a good predictive effect on pancreatic SPN invasiveness.The radiomics label based on pancreatic parenchyma phase images was the most predictive in the validation group and all 112 pancreatic SPNs.2.In the conventional CT model of the training group and the combined model with clinical data,the tumor capsule is an independent risk factor for predicting the invasiveness of pancreatic SPN;while in the nomogram model,the radiomics label and tumor capsule are independent risk factor for predicting the invasiveness of pancreatic SPN3.In multi-model comparison,the predictive performance of the nomogram model is higher than the clinical,conventional CT features and the combined model,and the Delong test shows that the difference is statistically significant.The nomogram model has higher prediction performance in both the training group and the verification group,and the Hosmer-Lemeshoe goodness-of-fit test indicates that the model has no deviation from the training group.Decision curve analysis shows that the nomogram model has high clinical value.更多还原